18 KiB
作成
LocalTicketBackend によって作成されました。
Intake summary
Panel から Quit するときに断続的な遅延が発生する問題について、調査・修正用 Ticket を作成した。現象は明確だが再現条件や遅延箇所は未特定のため readiness は spike_needed。Orchestrator はまず crates/tui/src/multi_pod.rs の Panel event loop / PendingReload / Quit handling / queue-attention notice dispatch / snapshot reload 周辺を調査し、Quit が非本質的な background work の完了待ちでブロックされない修正へ routing できる。関連 closed Ticket 00001KTFMMZP0 は Panel 非同期遷移改善で、今回の Quit 遅延とは別の follow-up として扱う。
State changed
要件・受け入れ条件・binding decisions・調査焦点が揃っており、Orchestrator が spike/実装修正へ routing できるため ready にする。
State changed
Ticket を workspace-panel が queued にしました。
Decision
Routing decision: spike_needed / queued_waiting_capacity
Reason:
- Ticket は
queuedで要件は明確だが、readiness はspike_needed。遅延原因・再現条件・影響 async path は未特定で、まずcrates/tui/src/multi_pod.rsの Panel event loop /PendingReload/ Quit handling / queue-attention notice dispatch / snapshot reload 周辺の code-path analysis が必要。 TicketRelationQueryとTicketOrchestrationPlanQueryで durable blocker はなかったが、この queue review では00001KV072V89(Workspace panel mouse selection)を先に inprogress 受理した。00001KV0723PCと00001KV072V89はどちらも Panel /multi_pod.rsの event/input/render/reload surface に触れる可能性が高く、同時 child worktree で進めると conflict risk と reviewer bottleneck が大きい。00001KV04NJ8Dは single-Pod rewind / Pod protocol 周辺の別 surface として parallel start 可能なため、現在の capacity は00001KV072V89と00001KV04NJ8Dに割り当てる。
Evidence checked:
- Ticket body / thread / artifacts(artifacts なし)。
- relation records: なし。
- orchestration plan records: この routing pass で
00001KV072V89とのconflicts_withと waiting-capacity note を記録。 - related closed Ticket
00001KTFMMZP0: Panel non-blocking transition work は closed で、PendingReload/ background refresh / attach-return path の背景として関連するが、今回の Quit 遅延とは別 follow-up。 - code map:
crates/tui/src/multi_pod.rsの run loop、PendingReload、Quit action、queue-attention notice dispatch、terminal event poll/reload interaction。 - workspace/Pod state: Orchestrator worktree clean。implementation Pods はまだ起動していないが、この routing pass で 2 件を受理予定。
Next action:
- この Ticket は queued のまま保持する。
00001KV072V89の branch が統合されるか、panel surface conflict が低いと再評価できた時点で、改めてqueued -> inprogress受理を検討する。- 受理時の first step は read-only/code-path spike とし、Quit が pending background work / notice dispatch / snapshot reload で block される path を特定してから修正する。
Escalate if:
- Quit 遅延の修正が terminal cleanup、Pod lifecycle authority、Ticket workflow semantics、または broad TUI runtime-loop redesign を必要とする場合。
- 原因が Panel 外(OS terminal / shell / external command / specific provider/network)にある証拠が出た場合。
Decision
Queue attention reviewed: no duplicate start.
Actual state check:
00001KV072V89isinprogresswith live spawned childcoder-00001KV072V89-panel-mouseand worktree/home/hare/Projects/yoi/.worktree/panel-mouse-selection.00001KV04NJ8Disinprogresswith live spawned childcoder-00001KV04NJ8D-rewindand worktree/home/hare/Projects/yoi/.worktree/rewind-live-refresh.- This Ticket remains
queued. TicketRelationQuerystill shows no durable dependency blockers, but the existing OrchestrationPlanconflicts_with/waiting_capacity_noteremains applicable: this Ticket likely touches the same Panelmulti_pod.rsevent-loop surface as00001KV072V89and should wait until that branch is integrated or conflict surface is re-evaluated.
Decision:
- Do not start another Coder Pod for this Ticket now.
- Keep queued; re-evaluate after
00001KV072V89completion/integration or explicit human override.
Decision
Routing decision: implementation_ready
Reason:
- Ticket は
queuedで、Quit 遅延の原因特定と修正の intent / requirements / acceptance criteria / binding invariants / implementation latitude / escalation conditions が揃っている。 readiness: spike_neededだが、調査対象はcrates/tui/src/multi_pod.rsの Panel event loop /PendingReload/ Quit handling / queue-attention notice dispatch / snapshot reload 周辺に bounded されており、専用 worktree の Coder に first-step code-path spike と修正を委ねられる。TicketRelationQueryに durable dependency blocker はない。- 以前の OrchestrationPlan
conflicts_with/waiting_capacity_noteは00001KV072V89(Panel mouse selection)との同時編集 conflict を理由に queued 待機としていたが、同 Ticket は review/merge/validation/cleanup 済みでdone。現在の Orchestrator worktree は clean で、related implementation Pods/worktrees も cleanup 済み。 - risk flags は
tui-panel/shutdown-latency/async-cancellationだが、Ticket は Quit の authority/invariants と escalation conditions を明記しており、実装前に追加の human decision は不要。
Evidence checked:
- Ticket body / thread / artifacts。
- relation records: なし。
- orchestration plan records: 既存
conflicts_with/waiting_capacity_noteを確認し、blocking condition(00001KV072V89の未統合)は解消済み。 - related Ticket
00001KTFMMZP0: Panel non-blocking transition work は closed で、PendingReload/ background refresh / attach-return path の背景として参照する。 - current code/workspace state:
00001KV072V89と00001KV04NJ8Dは merge/validation/cleanup 済み。Orchestrator worktree clean。visible spawned implementation Pods なし。
IntentPacket:
Intent:
yoi panel/ workspace Panel の Quit 操作で、ユーザー入力後に Panel の background reload / notice dispatch / snapshot load / Pod observation など非本質的処理待ちによる断続的な遅延が起きないよう、原因を特定して修正する。
Binding decisions / invariants:
- Quit はユーザーの明示的な終了意思であり、Panel の観測/reload/通知送信完了待ちで遅延してはならない。
- 端末状態復旧、描画 backend の正常終了、Rust drop による安全な abort など、最小限の終了処理は維持する。
- Ticket lifecycle authority、Pod authority boundary、Panel row/action semantics、Companion/Orchestrator composer target semantics は変更しない。
resources/promptsや durable Ticket schema 変更は想定しない。必要になった場合は escalation。
Requirements / acceptance criteria:
Ctrl+C/Ctrl+Dなど現行 Quit 経路が pending background work によって目に見えて遅延しない。- 遅延原因と修正方針を implementation report に説明する。
- 可能な範囲で unit test / regression test を追加し、Quit 経路が pending background work にブロックされないことを検証する。
- 自動化が難しい場合は manual validation 手順と観測結果を report する。
Implementation latitude:
- まず
crates/tui/src/multi_pod.rsの event loop /PendingReload/ Quit handling / terminal event polling / queue-attention notice dispatch / snapshot load を code-path spike する。 - 修正手段は、不要 await の回避、cancellable background work の abort/drop、event-loop priority adjustment、test seam 追加などから Coder が選んでよい。
- 再現が難しい場合は遅延し得る code path を単体再現できる test seam を優先してよい。
Escalate if:
- Panel public UX、Ticket lifecycle semantics、Pod shutdown semantics、authority boundary の変更が必要になる。
- 端末 cleanup や Pod process lifecycle safety を犠牲にしないと遅延を解消できない。
- 原因が Panel 外(OS terminal / shell / external command / provider/network)にある証拠が出た。
Validation:
- focused test for Quit not waiting on pending background work。
cargo fmt --check。git diff --check。- 変更範囲に応じて
cargo test -p tui .../cargo check -p tui --all-targets。 - 可能なら manual
yoi panelQuit observation を report。
Current code map:
crates/tui/src/multi_pod.rs: Panel run loop、Quit key handling、PendingReload、background reload / attention notice / snapshot observation paths。crates/tui/src/workspace_panel.rs: ViewModel は参照のみ。row/action semantics は変更しない。
Critical risks / reviewer focus:
- Quit を速くするために terminal cleanup / safe abort / Pod lifecycle を壊していないこと。
- nonessential background work が Quit を block しないこと。
- Panel/Ticket/Pod authority semantics と row/action semantics が変わっていないこと。
- Regression test が実際に pending background work block を防ぐ性質を持つこと。
State changed
Ticket evidence、relation records、existing OrchestrationPlan waiting/conflict notes、related completed Panel mouse integration、visible Pod/worktree state、and clean Orchestrator workspace were checked. The prior conflict/capacity blocker is resolved; no durable dependency blockers remain. Implementation can proceed in a dedicated child worktree with a first-step code-path spike and focused regression test before changing Quit behavior.
Implementation report
Implementation report for Ticket 00001KV0723PC
Commit: cfe411e50d (fix: avoid panel quit notice wait)
Observed / identified delay cause:
- In
crates/tui/src/multi_pod.rs, the workspace Panel run loop handled a completedPendingReloadbefore polling terminal input. - When the reload result requested Orchestrator queue-attention notification, the loop awaited
dispatch_orchestrator_queue_attention_notice(request)inline. - That dispatch path can wait on Unix socket connect/read/write operations and their timeouts. While it was awaited inline, queued terminal events such as
Ctrl+C/Ctrl+Dcould not be observed, so an explicit Quit could be delayed by non-essential notice dispatch work.
Fix summary:
- Added
PendingQueueAttentionNotice, a cancellable background task wrapper for queue-attention notice dispatch. - Changed the Panel run loop to start queue-attention notice dispatch asynchronously after reload and only harvest the result when the task is already finished.
- Added explicit quit cleanup through
abort_panel_background_work_for_quit, aborting both pending reload and pending queue-attention notice work before returningMultiPodOutcome::Quit. - Also abort the background notice before foreground open/action paths that already abort reload work, so non-essential notice dispatch does not contend with user-directed operations.
- Preserved terminal/backend shutdown behavior by keeping the existing return path and relying on explicit abort plus Rust
Dropcleanup for background tasks.
Files changed:
crates/tui/src/multi_pod.rs
Regression coverage:
- Added
multi_quit_aborts_background_reload_and_notice_without_waiting, which starts never-completing reload and notice tasks, exercises the quit abort helper under a timeout, and verifies both tasks are cancelled. - Existing reload overlap behavior remains covered by
multi_poll_reload_does_not_overlap_in_flight_reload.
Validation commands/results:
cargo fmt --check— passedgit diff --check— passedcargo test -p tui multi_quit_aborts_background_reload_and_notice_without_waiting --lib— passedcargo test -p tui multi_poll_reload_does_not_overlap_in_flight_reload --lib— passedcargo check -p tui --all-targets— passed
Residual risks / manual validation notes:
- No broad runtime-loop rewrite was done. Foreground user-directed operations that are already awaited by the Panel can still occupy the loop; this change targets the identified non-essential background notice dispatch delay path.
- I did not run an interactive manual
yoi panelsession from this Coder environment. Recommended manual check: startyoi panel, trigger a reload/queued Ticket condition that would dispatch queue-attention notice, then pressCtrl+C/Ctrl+Dwhile the notice target is slow or unavailable; the Panel should exit promptly while terminal state is restored.
Review: approve
Review: approve
Evidence:
- Reviewed Ticket intent/criteria and
git diff 9dacc90e..HEADfor commitscfe411e5and8de82836. - The reported delay cause is plausible in the previous loop shape: a completed reload was processed before input polling and could inline-await queue-attention notice dispatch; that dispatch uses Pod socket connect/read/write paths, so queued
Ctrl+C/Ctrl+Dinput could wait behind nonessential notice work. - The implementation changes only
crates/tui/src/multi_pod.rsplus the implementation report. Queue-attention notice dispatch is now wrapped in a cancellablePendingQueueAttentionNotice, started in the background, and harvested only after completion. Reload dispatch is guarded against overlap. - Quit handling now calls
abort_panel_background_work_for_quit, aborting pending reload and pending queue-attention notice tasks before returning the existingMultiPodOutcome::Quit; this keeps the established terminal/backend cleanup return path intact. - Existing foreground open/action helper paths also abort the nonessential pending notice before returning user-directed outcomes, so stale notice dispatch does not block those paths.
- No Ticket lifecycle authority, Pod authority boundary, Panel row/action semantics, composer target semantics, prompt resources, or Ticket schema changes were introduced.
- The regression test starts never-completing reload and notice futures, exercises the quit abort helper under a timeout, and verifies cancellation of both tasks. This is a meaningful regression for the pending-background-work latency path rather than only checking that a helper was wired.
Validation performed:
cargo test -p tui multi_quit_aborts_background_reload_and_notice_without_waiting --lib— passed.cargo test -p tui multi_poll_reload_does_not_overlap_in_flight_reload --lib— passed.cargo fmt --check— passed.git diff --check— passed.cargo check -p tui --all-targets— passed.
Risks / notes:
- I did not run an interactive
yoi panelmanual latency check. The code-path review and focused tests cover the identified nonessential reload/notice wait path.
Implementation report
Integration outcome: merged and validated.
- Implementation branch:
ticket-00001KV0723PC-panel-quit-latency - Implementation commits:
cfe411e5 fix: avoid panel quit notice wait,8de82836 ticket: report panel quit latency fix - Review commit:
2bb36cc4 ticket: approve panel quit latency - Orchestrator merge commit:
db7bad7a merge: panel quit latency
Reviewer result:
approve。旧経路で reload 完了後に queue-attention notice dispatch を inline await していたことが Quit 遅延原因として妥当であること、実装が background/cancellable 化と Quit abort cleanup で解消していること、terminal/backend cleanup path と authority semantics を維持していることを確認済み。
Orchestrator validation after merge:
cargo test -p tui multi_quit_aborts_background_reload_and_notice_without_waiting --lib: PASScargo test -p tui multi_poll_reload_does_not_overlap_in_flight_reload --lib: PASScargo fmt --check: PASSgit diff --check: PASScargo check -p tui --all-targets: PASS
Residual notes:
- Interactive
yoi panelmanual latency check は未実施。ただし、code-path review と focused tests で特定された nonessential reload/notice wait path は cover されている。 - Ticket lifecycle authority、Pod authority boundary、Panel row/action semantics、composer target semantics、prompt resources、Ticket schema は変更していない。
Next:
- Ticket を
doneに遷移し、panel quit child Pods / worktree / branch を cleanup する。close は別途明示権限がある場合に行う。
State changed
Implementation branch ticket-00001KV0723PC-panel-quit-latency was reviewed, approved, merged into the Orchestrator branch as db7bad7a, and validated in the Orchestrator worktree. Focused panel quit latency tests, formatting, diff check, and cargo check -p tui --all-targets passed. Ticket implementation work is done; closure remains separate.